Grain is not an overlay.
Grain is the image.
Our goal is to reconstruct how exposure, finite silver-halide sites, dye clouds, interlayer inhibition, printing and scanning form one image.
Match the frame before judging 35 mm grain and sharpness
V36 does not make the grain finer or hide it with extra softness. The audit found that V35 T007 and T031 began at frame 0, while V34 used the curated frame-276 and frame-132 windows. Source motion and texture were therefore presented as a film-model change. V36 locks camera, projection, scan, still and hover video to the same absolute source frames, then rechecks processed-stock MTF and 48 μm diffuse RMS granularity on one physical scale. At the correct T031 window, Production differs from V34 by only about 0.1% in high-frequency, temporal and grain-to-edge measures, so the film model stays frozen.
Noise is not an overlay.
Randomness participates in image formation.
Digital noise is usually added after the picture exists. Colour-negative randomness happens before the image is complete: whether finite silver-halide sites develop changes dye amount, local density, DIR release and neighbouring layers together.
After development, silver is bleached and fixed away; printing and scanning observe dye clouds. We therefore simulate a dye-cloud field seeded by silver-halide events, not permanent silver specks floating over RGB.
Baseline principle: reconstruct film, printing and scanning; keep creative grading outside the model.
See equations and key code →Every physical effect gets one explicit place in the image-formation chain.
- Match the source frameV36 locks T002 0–23, T007 276–299 and T031 132–155; camera, projection, scan, still and video must agree.
- One encode per masterThe V31 chroma boundary runs in linear Rec.709 memory; neither observer takes an intermediate ProRes round trip.
- Unknown means frozenWithout 5279 NPS or separation-wedge measurements, grain morphology, interimage DIR and green neutralization stay frozen.
- Auditable stochasticityAll 45 Philox-u32 site identities per frame must be unique; unsafe concurrency and rare 2383 threshold outliers block release.
Version archive
From V4 to V36, compare both viewing chains and retain every error and correction.
OPEN →PAPERResearch paper
Emulsion layers, masking, DIR, grain, scanners and the ten-note evidence audit.
OPEN →METHODMethod
Equations, flow and key code from RAW through the negative, 2383 and Cineon.
OPEN →